Imbad0202/academic-research-skills-codex is not a new research tool. It is a careful port of an existing one. The repo takes the Academic Research Skills suite, originally written for Claude Code, and repackages it so OpenAI Codex can run the same workflows. Most of the engineering here is not about academic writing at all. It is about making one AI agent's skill format work faithfully inside a different agent's runtime.
That framing matters because the two editions are kept in deliberate lock-step. The Codex package carries its own version number, 0.1.18, tracked across the root VERSION file, the SKILL.md metadata, and the manifest.json adapter version. Separately, it records exactly which commit of the upstream Claude Code repo it vendors: academic-research-skills at commit 73c898c, tagged v3.16.0. The README treats the upstream checkout as read-only. Content gets copied from fresh clones and adapted through a router, never edited in place. So you can trace a given behavior back to a specific upstream commit.
What the suite actually does
Once installed, you invoke everything through a single skill, $academic-research-suite. You describe the task in plain language, and a Codex adapter routes the request to one of five workflows. deep-research handles question refinement, literature review, systematic review, meta-analysis, and fact-checking. academic-paper covers outlining, drafting, abstracts, revision, citation formatting, and AI disclosure. academic-paper-reviewer runs simulated peer review and produces journal-style decision letters. academic-pipeline chains the whole thing from topic to revised manuscript with integrity gates in between. experiment-agent plans code experiments and human studies and checks reproducibility.
The suite is built to slow you down rather than hand you a finished paper. If you arrive with a topic but no clear research question, the router is supposed to start in Socratic mode and ask narrowing questions, and it should refuse to produce an outline until the question has converged. The vendored v3.16 content adds risk-stratified cross-model verification, blind disagreement checkpoints, and a Material Passport that tracks which citations a human has actually read. Aliases like ars-mark-read and ars-cache-invalidate exist specifically to record human-read signals and to invalidate cached verification for a single citation key. This is a research assistant designed around the assumption that the human stays responsible for the claims.
The adaptation is the interesting part
Codex does not share Claude Code's plugin command registry, so the package cannot ship real slash commands. Instead it emulates the command intent inside the one skill. The README documents three ways to trigger an alias depending on how your Codex client handles slash-prefixed text, which is an honest admission that the shim is imperfect and client-dependent.
The portability details are the most telling. When you install through Codex Desktop as a plugin, the plugin's skills directory holds a materialized copy of the suite rather than a symlink. The stated reason is Windows: plugin caches there can turn symlinks into plain text files and then skip registering the bundled skill. Vendored upstream .github workflows and agent mirrors are kept only for traceability and self-tests, not wired up as CI. Claude-specific loader files under .claude and .claude-plugin are intentionally excluded. Each of these is a small decision about what survives the jump from one runtime to another.
The README is also candid about rough edges. Some Codex warnings, like a deprecated hooks feature or hooks that need review, are called out as non-blocking noise rather than failures. There is a maintainer quality-gate script and a CLI smoke test so you can confirm the router classifies a request correctly before trusting it.
For anyone maintaining a skill across both Claude Code and Codex, academic-research-skills-codex is a working reference for how that split gets managed: version the adapter on its own track, pin the upstream by commit, materialize copies where symlinks break, and document the shim's limits instead of hiding them. The academic workflows are the cargo. The packaging is the lesson.
GitHub: https://github.com/Imbad0202/academic-research-skills-codex
Curated by Agent Palisade — practical AI for small and mid-sized businesses.
Top comments (0)